Simultaneous Bayesian clustering and feature selection using RJMCMC-based learning of finite generalized Dirichlet mixture models

Selecting relevant features in multidimensional data is important in several pattern analysis and image processing applications. The goal of this paper is to propose a Bayesian approach for identifying clusters of proportional data based on the selection of relevant features. More specifically, we consider the problem of selecting relevant features in unsupervised settings when generalized Dirichlet mixture models are considered to model and cluster proportional data. The learning of the proposed statistical model, to formulate the unsupervised feature selection problem, is carried out using a powerful reversible jump Markov chain Monte Carlo (RJMCMC) technique. Experiments involving the challenging problems of human action videos categorization, pedestrian detection and face recognition indicate that the proposed approach is efficient.

► A simultaneous approach for clustering and feature selection is presented. ► The proposed approach is based on the generalized Dirichlet mixture and learned in a Bayesian way. ► A complete RJMCMC approach is developed. ► The proposed approach is applied to human action videos categorization, pedestrian detection and face recognition.